Imagine that you have several emotions as predictors (input layers) of a certain output (lets say... wellbeing). You create an ANN and the best architecture contains 5 hidden nodes.
To determine the contribution of the individual emotions, apart from sensitivity analysis, can we also measure the value of the weights of the connections between them and the hidden nodes?
And if the contribution of a given emotion to the most important hidden nodes is inhibitory, can we assume that the relation of that emotion and the outcome is negative?
TIA
I don't think it is that simple.
You should also observe weights between hidden nodes and an output node. It is possible, that weight between input and a certain hidden node is negative, but so does weight between this node and output of the network. Then - the higher the input with negative weight, the higher the final output returned by the network.
What is more, If one of the hidden nodes have some weights in negative, it might also mean, that remaining inputs can contribute to output positively only if they are higher than certain input (With negative weight).
With regards,
Ziemowit
Hi Rita,
Ziemowit is right - it's not that simple. Check my paper available from ResearchGate:
https://www.researchgate.net/publication/215610896_Busting_the_Black_Box_Myth_Designing_Out_Unwanted_ADMET_Properties_with_Machine_Learning_Approaches
It reports another method that you may find useful in answering your question.
Article Busting the Black Box Myth: Designing Out Unwanted ADMET Pro...
Dear Rita,
There is a real number associated with each connection in ANN, which is called the weight of the connection. We denote by Wij the weight of the connection from unit ui to unit uj. It is then convenient to represent the pattern of connectivity in the network by a weight matrix W whose elements are the weights Wij. Two types of connection are usually distinguished: excitatory and inhibitory. A positive weight represents an excitatory connection whereas a negative weight represents an inhibitory connection.
This is the general explanation of negative weights others explains it as the neurons or "nodes" of an ANN correspond to the excitatory neurons of the brain. The anti-correlations represented by the negative weights may instead be represented by parallel opposing populations of excitatory neurons which are kept separated and anti-correlated by inhibitory neurons that connect between them.
So it is hard to say that having negative values means negative emotions. I better try different ANN architecture.
What ANN architecture are you using?
- Badges
- Science topic